2 * Copyright (c) 2006 QLogic, Inc. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include "ipath_verbs.h"
35 #include "ipath_kernel.h"
37 /* cut down ridiculously long IB macro names */
38 #define OP(x) IB_OPCODE_RC_##x
40 static u32 restart_sge(struct ipath_sge_state *ss, struct ipath_swqe *wqe,
45 len = ((psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
46 ss->sge = wqe->sg_list[0];
47 ss->sg_list = wqe->sg_list + 1;
48 ss->num_sge = wqe->wr.num_sge;
49 ipath_skip_sge(ss, len);
50 return wqe->length - len;
54 * ipath_init_restart- initialize the qp->s_sge after a restart
55 * @qp: the QP who's SGE we're restarting
56 * @wqe: the work queue to initialize the QP's SGE from
58 * The QP s_lock should be held and interrupts disabled.
60 static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
62 struct ipath_ibdev *dev;
64 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn,
65 ib_mtu_enum_to_int(qp->path_mtu));
66 dev = to_idev(qp->ibqp.device);
67 spin_lock(&dev->pending_lock);
68 if (list_empty(&qp->timerwait))
69 list_add_tail(&qp->timerwait,
70 &dev->pending[dev->pending_index]);
71 spin_unlock(&dev->pending_lock);
75 * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
76 * @qp: a pointer to the QP
77 * @ohdr: a pointer to the IB header being constructed
80 * Return 1 if constructed; otherwise, return 0.
81 * Note that we are in the responder's side of the QP context.
82 * Note the QP s_lock must be held.
84 static int ipath_make_rc_ack(struct ipath_qp *qp,
85 struct ipath_other_headers *ohdr,
86 u32 pmtu, u32 *bth0p, u32 *bth2p)
88 struct ipath_ack_entry *e;
94 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
97 switch (qp->s_ack_state) {
98 case OP(RDMA_READ_RESPONSE_LAST):
99 case OP(RDMA_READ_RESPONSE_ONLY):
100 case OP(ATOMIC_ACKNOWLEDGE):
101 qp->s_ack_state = OP(ACKNOWLEDGE);
103 case OP(ACKNOWLEDGE):
104 /* Check for no next entry in the queue. */
105 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
106 if (qp->s_flags & IPATH_S_ACK_PENDING)
111 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
112 if (e->opcode == OP(RDMA_READ_REQUEST)) {
113 /* Copy SGE state in case we need to resend */
114 qp->s_ack_rdma_sge = e->rdma_sge;
115 qp->s_cur_sge = &qp->s_ack_rdma_sge;
116 len = e->rdma_sge.sge.sge_length;
119 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
121 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
122 if (++qp->s_tail_ack_queue >
123 IPATH_MAX_RDMA_ATOMIC)
124 qp->s_tail_ack_queue = 0;
126 ohdr->u.aeth = ipath_compute_aeth(qp);
128 qp->s_ack_rdma_psn = e->psn;
129 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
131 /* COMPARE_SWAP or FETCH_ADD */
132 qp->s_cur_sge = NULL;
134 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
135 ohdr->u.at.aeth = ipath_compute_aeth(qp);
136 ohdr->u.at.atomic_ack_eth[0] =
137 cpu_to_be32(e->atomic_data >> 32);
138 ohdr->u.at.atomic_ack_eth[1] =
139 cpu_to_be32(e->atomic_data);
140 hwords += sizeof(ohdr->u.at) / sizeof(u32);
142 if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
143 qp->s_tail_ack_queue = 0;
145 bth0 = qp->s_ack_state << 24;
148 case OP(RDMA_READ_RESPONSE_FIRST):
149 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
151 case OP(RDMA_READ_RESPONSE_MIDDLE):
152 len = qp->s_ack_rdma_sge.sge.sge_length;
156 ohdr->u.aeth = ipath_compute_aeth(qp);
158 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
159 if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
160 qp->s_tail_ack_queue = 0;
162 bth0 = qp->s_ack_state << 24;
163 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
169 * Send a regular ACK.
170 * Set the s_ack_state so we wait until after sending
171 * the ACK before setting s_ack_state to ACKNOWLEDGE
174 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
175 qp->s_flags &= ~IPATH_S_ACK_PENDING;
176 qp->s_cur_sge = NULL;
179 cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
181 IPATH_AETH_CREDIT_SHIFT));
183 ohdr->u.aeth = ipath_compute_aeth(qp);
186 bth0 = OP(ACKNOWLEDGE) << 24;
187 bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
189 qp->s_hdrwords = hwords;
190 qp->s_cur_size = len;
200 * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
201 * @qp: a pointer to the QP
202 * @ohdr: a pointer to the IB header being constructed
203 * @pmtu: the path MTU
204 * @bth0p: pointer to the BTH opcode word
205 * @bth2p: pointer to the BTH PSN word
207 * Return 1 if constructed; otherwise, return 0.
208 * Note the QP s_lock must be held and interrupts disabled.
210 int ipath_make_rc_req(struct ipath_qp *qp,
211 struct ipath_other_headers *ohdr,
212 u32 pmtu, u32 *bth0p, u32 *bth2p)
214 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
215 struct ipath_sge_state *ss;
216 struct ipath_swqe *wqe;
223 /* Sending responses has higher priority over sending requests. */
224 if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
225 (qp->s_flags & IPATH_S_ACK_PENDING) ||
226 qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE) &&
227 ipath_make_rc_ack(qp, ohdr, pmtu, bth0p, bth2p))
230 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
231 qp->s_rnr_timeout || qp->s_wait_credit)
234 /* Limit the number of packets sent without an ACK. */
235 if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) {
236 qp->s_wait_credit = 1;
241 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
245 /* Send a request. */
246 wqe = get_swqe_ptr(qp, qp->s_cur);
247 switch (qp->s_state) {
250 * Resend an old request or start a new one.
252 * We keep track of the current SWQE so that
253 * we don't reset the "furthest progress" state
254 * if we need to back up.
257 if (qp->s_cur == qp->s_tail) {
258 /* Check if send work queue is empty. */
259 if (qp->s_tail == qp->s_head)
262 * If a fence is requested, wait for previous
263 * RDMA read and atomic operations to finish.
265 if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
266 qp->s_num_rd_atomic) {
267 qp->s_flags |= IPATH_S_FENCE_PENDING;
270 wqe->psn = qp->s_next_psn;
274 * Note that we have to be careful not to modify the
275 * original work request since we may need to resend
281 switch (wqe->wr.opcode) {
283 case IB_WR_SEND_WITH_IMM:
284 /* If no credit, return. */
285 if (qp->s_lsn != (u32) -1 &&
286 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
288 wqe->lpsn = wqe->psn;
290 wqe->lpsn += (len - 1) / pmtu;
291 qp->s_state = OP(SEND_FIRST);
295 if (wqe->wr.opcode == IB_WR_SEND)
296 qp->s_state = OP(SEND_ONLY);
298 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
299 /* Immediate data comes after the BTH */
300 ohdr->u.imm_data = wqe->wr.imm_data;
303 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
305 bth2 = 1 << 31; /* Request ACK. */
306 if (++qp->s_cur == qp->s_size)
310 case IB_WR_RDMA_WRITE:
311 if (newreq && qp->s_lsn != (u32) -1)
314 case IB_WR_RDMA_WRITE_WITH_IMM:
315 /* If no credit, return. */
316 if (qp->s_lsn != (u32) -1 &&
317 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
319 ohdr->u.rc.reth.vaddr =
320 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
321 ohdr->u.rc.reth.rkey =
322 cpu_to_be32(wqe->wr.wr.rdma.rkey);
323 ohdr->u.rc.reth.length = cpu_to_be32(len);
324 hwords += sizeof(struct ib_reth) / sizeof(u32);
325 wqe->lpsn = wqe->psn;
327 wqe->lpsn += (len - 1) / pmtu;
328 qp->s_state = OP(RDMA_WRITE_FIRST);
332 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
333 qp->s_state = OP(RDMA_WRITE_ONLY);
336 OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
337 /* Immediate data comes after RETH */
338 ohdr->u.rc.imm_data = wqe->wr.imm_data;
340 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
343 bth2 = 1 << 31; /* Request ACK. */
344 if (++qp->s_cur == qp->s_size)
348 case IB_WR_RDMA_READ:
350 * Don't allow more operations to be started
351 * than the QP limits allow.
354 if (qp->s_num_rd_atomic >=
355 qp->s_max_rd_atomic) {
356 qp->s_flags |= IPATH_S_RDMAR_PENDING;
359 qp->s_num_rd_atomic++;
360 if (qp->s_lsn != (u32) -1)
363 * Adjust s_next_psn to count the
364 * expected number of responses.
367 qp->s_next_psn += (len - 1) / pmtu;
368 wqe->lpsn = qp->s_next_psn++;
370 ohdr->u.rc.reth.vaddr =
371 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
372 ohdr->u.rc.reth.rkey =
373 cpu_to_be32(wqe->wr.wr.rdma.rkey);
374 ohdr->u.rc.reth.length = cpu_to_be32(len);
375 qp->s_state = OP(RDMA_READ_REQUEST);
376 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
379 if (++qp->s_cur == qp->s_size)
383 case IB_WR_ATOMIC_CMP_AND_SWP:
384 case IB_WR_ATOMIC_FETCH_AND_ADD:
386 * Don't allow more operations to be started
387 * than the QP limits allow.
390 if (qp->s_num_rd_atomic >=
391 qp->s_max_rd_atomic) {
392 qp->s_flags |= IPATH_S_RDMAR_PENDING;
395 qp->s_num_rd_atomic++;
396 if (qp->s_lsn != (u32) -1)
398 wqe->lpsn = wqe->psn;
400 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
401 qp->s_state = OP(COMPARE_SWAP);
402 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
403 wqe->wr.wr.atomic.swap);
404 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
405 wqe->wr.wr.atomic.compare_add);
407 qp->s_state = OP(FETCH_ADD);
408 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
409 wqe->wr.wr.atomic.compare_add);
410 ohdr->u.atomic_eth.compare_data = 0;
412 ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
413 wqe->wr.wr.atomic.remote_addr >> 32);
414 ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
415 wqe->wr.wr.atomic.remote_addr);
416 ohdr->u.atomic_eth.rkey = cpu_to_be32(
417 wqe->wr.wr.atomic.rkey);
418 hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
421 if (++qp->s_cur == qp->s_size)
428 qp->s_sge.sge = wqe->sg_list[0];
429 qp->s_sge.sg_list = wqe->sg_list + 1;
430 qp->s_sge.num_sge = wqe->wr.num_sge;
431 qp->s_len = wqe->length;
434 if (qp->s_tail >= qp->s_size)
437 bth2 |= qp->s_psn & IPATH_PSN_MASK;
438 if (wqe->wr.opcode == IB_WR_RDMA_READ)
439 qp->s_psn = wqe->lpsn + 1;
442 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
443 qp->s_next_psn = qp->s_psn;
446 * Put the QP on the pending list so lost ACKs will cause
447 * a retry. More than one request can be pending so the
448 * QP may already be on the dev->pending list.
450 spin_lock(&dev->pending_lock);
451 if (list_empty(&qp->timerwait))
452 list_add_tail(&qp->timerwait,
453 &dev->pending[dev->pending_index]);
454 spin_unlock(&dev->pending_lock);
457 case OP(RDMA_READ_RESPONSE_FIRST):
459 * This case can only happen if a send is restarted.
460 * See ipath_restart_rc().
462 ipath_init_restart(qp, wqe);
465 qp->s_state = OP(SEND_MIDDLE);
467 case OP(SEND_MIDDLE):
468 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
469 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
470 qp->s_next_psn = qp->s_psn;
477 if (wqe->wr.opcode == IB_WR_SEND)
478 qp->s_state = OP(SEND_LAST);
480 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
481 /* Immediate data comes after the BTH */
482 ohdr->u.imm_data = wqe->wr.imm_data;
485 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
487 bth2 |= 1 << 31; /* Request ACK. */
489 if (qp->s_cur >= qp->s_size)
493 case OP(RDMA_READ_RESPONSE_LAST):
495 * This case can only happen if a RDMA write is restarted.
496 * See ipath_restart_rc().
498 ipath_init_restart(qp, wqe);
500 case OP(RDMA_WRITE_FIRST):
501 qp->s_state = OP(RDMA_WRITE_MIDDLE);
503 case OP(RDMA_WRITE_MIDDLE):
504 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
505 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
506 qp->s_next_psn = qp->s_psn;
513 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
514 qp->s_state = OP(RDMA_WRITE_LAST);
516 qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
517 /* Immediate data comes after the BTH */
518 ohdr->u.imm_data = wqe->wr.imm_data;
520 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
523 bth2 |= 1 << 31; /* Request ACK. */
525 if (qp->s_cur >= qp->s_size)
529 case OP(RDMA_READ_RESPONSE_MIDDLE):
531 * This case can only happen if a RDMA read is restarted.
532 * See ipath_restart_rc().
534 ipath_init_restart(qp, wqe);
535 len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
536 ohdr->u.rc.reth.vaddr =
537 cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
538 ohdr->u.rc.reth.rkey =
539 cpu_to_be32(wqe->wr.wr.rdma.rkey);
540 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
541 qp->s_state = OP(RDMA_READ_REQUEST);
542 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
543 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
544 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
545 qp->s_next_psn = qp->s_psn;
549 if (qp->s_cur == qp->s_size)
553 if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
554 bth2 |= 1 << 31; /* Request ACK. */
556 qp->s_hdrwords = hwords;
558 qp->s_cur_size = len;
559 *bth0p = bth0 | (qp->s_state << 24);
569 * send_rc_ack - Construct an ACK packet and send it
570 * @qp: a pointer to the QP
572 * This is called from ipath_rc_rcv() and only uses the receive
574 * Note that RDMA reads and atomics are handled in the
575 * send side QP state and tasklet.
577 static void send_rc_ack(struct ipath_qp *qp)
579 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
583 struct ipath_ib_header hdr;
584 struct ipath_other_headers *ohdr;
587 /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
588 if (qp->r_head_ack_queue != qp->s_tail_ack_queue)
591 /* Construct the header. */
593 lrh0 = IPATH_LRH_BTH;
594 /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
596 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
597 hwords += ipath_make_grh(dev, &hdr.u.l.grh,
598 &qp->remote_ah_attr.grh,
601 lrh0 = IPATH_LRH_GRH;
603 /* read pkey_index w/o lock (its atomic) */
604 bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
605 OP(ACKNOWLEDGE) << 24;
607 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
609 IPATH_AETH_CREDIT_SHIFT));
611 ohdr->u.aeth = ipath_compute_aeth(qp);
612 lrh0 |= qp->remote_ah_attr.sl << 4;
613 hdr.lrh[0] = cpu_to_be16(lrh0);
614 hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
615 hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
616 hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
617 ohdr->bth[0] = cpu_to_be32(bth0);
618 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
619 ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
622 * If we can send the ACK, clear the ACK state.
624 if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) {
625 dev->n_unicast_xmit++;
630 * We are out of PIO buffers at the moment.
631 * Pass responsibility for sending the ACK to the
632 * send tasklet so that when a PIO buffer becomes
633 * available, the ACK is sent ahead of other outgoing
639 spin_lock_irqsave(&qp->s_lock, flags);
640 qp->s_flags |= IPATH_S_ACK_PENDING;
641 qp->s_nak_state = qp->r_nak_state;
642 qp->s_ack_psn = qp->r_ack_psn;
643 spin_unlock_irqrestore(&qp->s_lock, flags);
645 /* Call ipath_do_rc_send() in another thread. */
646 tasklet_hi_schedule(&qp->s_task);
653 * reset_psn - reset the QP state to send starting from PSN
655 * @psn: the packet sequence number to restart at
657 * This is called from ipath_rc_rcv() to process an incoming RC ACK
659 * Called at interrupt level with the QP s_lock held.
661 static void reset_psn(struct ipath_qp *qp, u32 psn)
664 struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
670 * If we are starting the request from the beginning,
671 * let the normal send code handle initialization.
673 if (ipath_cmp24(psn, wqe->psn) <= 0) {
674 qp->s_state = OP(SEND_LAST);
678 /* Find the work request opcode corresponding to the given PSN. */
679 opcode = wqe->wr.opcode;
683 if (++n == qp->s_size)
687 wqe = get_swqe_ptr(qp, n);
688 diff = ipath_cmp24(psn, wqe->psn);
693 * If we are starting the request from the beginning,
694 * let the normal send code handle initialization.
697 qp->s_state = OP(SEND_LAST);
700 opcode = wqe->wr.opcode;
704 * Set the state to restart in the middle of a request.
705 * Don't change the s_sge, s_cur_sge, or s_cur_size.
706 * See ipath_do_rc_send().
710 case IB_WR_SEND_WITH_IMM:
711 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
714 case IB_WR_RDMA_WRITE:
715 case IB_WR_RDMA_WRITE_WITH_IMM:
716 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
719 case IB_WR_RDMA_READ:
720 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
725 * This case shouldn't happen since its only
728 qp->s_state = OP(SEND_LAST);
735 * ipath_restart_rc - back up requester to resend the last un-ACKed request
736 * @qp: the QP to restart
737 * @psn: packet sequence number for the request
738 * @wc: the work completion request
740 * The QP s_lock should be held and interrupts disabled.
742 void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
744 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
745 struct ipath_ibdev *dev;
747 if (qp->s_retry == 0) {
748 wc->wr_id = wqe->wr.wr_id;
749 wc->status = IB_WC_RETRY_EXC_ERR;
750 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
754 wc->src_qp = qp->remote_qpn;
756 wc->slid = qp->remote_ah_attr.dlid;
757 wc->sl = qp->remote_ah_attr.sl;
758 wc->dlid_path_bits = 0;
760 ipath_sqerror_qp(qp, wc);
766 * Remove the QP from the timeout queue.
767 * Note: it may already have been removed by ipath_ib_timer().
769 dev = to_idev(qp->ibqp.device);
770 spin_lock(&dev->pending_lock);
771 if (!list_empty(&qp->timerwait))
772 list_del_init(&qp->timerwait);
773 spin_unlock(&dev->pending_lock);
775 if (wqe->wr.opcode == IB_WR_RDMA_READ)
778 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
781 tasklet_hi_schedule(&qp->s_task);
787 static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
789 if (qp->s_wait_credit) {
790 qp->s_wait_credit = 0;
791 tasklet_hi_schedule(&qp->s_task);
793 qp->s_last_psn = psn;
797 * do_rc_ack - process an incoming RC ACK
798 * @qp: the QP the ACK came in on
799 * @psn: the packet sequence number of the ACK
800 * @opcode: the opcode of the request that resulted in the ACK
802 * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
804 * Called at interrupt level with the QP s_lock held and interrupts disabled.
805 * Returns 1 if OK, 0 if current operation should be aborted (NAK).
807 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
809 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
811 struct ipath_swqe *wqe;
816 * Remove the QP from the timeout queue (or RNR timeout queue).
817 * If ipath_ib_timer() has already removed it,
818 * it's OK since we hold the QP s_lock and ipath_restart_rc()
819 * just won't find anything to restart if we ACK everything.
821 spin_lock(&dev->pending_lock);
822 if (!list_empty(&qp->timerwait))
823 list_del_init(&qp->timerwait);
824 spin_unlock(&dev->pending_lock);
827 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
828 * requests and implicitly NAK RDMA read and atomic requests issued
829 * before the NAK'ed request. The MSN won't include the NAK'ed
830 * request but will include an ACK'ed request(s).
835 wqe = get_swqe_ptr(qp, qp->s_last);
838 * The MSN might be for a later WQE than the PSN indicates so
839 * only complete WQEs that the PSN finishes.
841 while (ipath_cmp24(ack_psn, wqe->lpsn) >= 0) {
843 * If this request is a RDMA read or atomic, and the ACK is
844 * for a later operation, this ACK NAKs the RDMA read or
845 * atomic. In other words, only a RDMA_READ_LAST or ONLY
846 * can ACK a RDMA read and likewise for atomic ops. Note
847 * that the NAK case can only happen if relaxed ordering is
848 * used and requests are sent after an RDMA read or atomic
849 * is sent but before the response is received.
851 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
852 (opcode != OP(RDMA_READ_RESPONSE_LAST) ||
853 ipath_cmp24(ack_psn, wqe->lpsn) != 0)) ||
854 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
855 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
856 (opcode != OP(ATOMIC_ACKNOWLEDGE) ||
857 ipath_cmp24(wqe->psn, psn) != 0))) {
859 * The last valid PSN seen is the previous
862 update_last_psn(qp, wqe->psn - 1);
863 /* Retry this request. */
864 ipath_restart_rc(qp, wqe->psn, &wc);
866 * No need to process the ACK/NAK since we are
867 * restarting an earlier request.
871 if (qp->s_num_rd_atomic &&
872 (wqe->wr.opcode == IB_WR_RDMA_READ ||
873 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
874 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
875 qp->s_num_rd_atomic--;
876 /* Restart sending task if fence is complete */
877 if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
878 !qp->s_num_rd_atomic) {
879 qp->s_flags &= ~IPATH_S_FENCE_PENDING;
880 tasklet_hi_schedule(&qp->s_task);
881 } else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
882 qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
883 tasklet_hi_schedule(&qp->s_task);
886 /* Post a send completion queue entry if requested. */
887 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
888 (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
889 wc.wr_id = wqe->wr.wr_id;
890 wc.status = IB_WC_SUCCESS;
891 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
893 wc.byte_len = wqe->length;
896 wc.src_qp = qp->remote_qpn;
899 wc.slid = qp->remote_ah_attr.dlid;
900 wc.sl = qp->remote_ah_attr.sl;
901 wc.dlid_path_bits = 0;
903 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
905 qp->s_retry = qp->s_retry_cnt;
907 * If we are completing a request which is in the process of
908 * being resent, we can stop resending it since we know the
909 * responder has already seen it.
911 if (qp->s_last == qp->s_cur) {
912 if (++qp->s_cur >= qp->s_size)
914 qp->s_last = qp->s_cur;
915 if (qp->s_last == qp->s_tail)
917 wqe = get_swqe_ptr(qp, qp->s_cur);
918 qp->s_state = OP(SEND_LAST);
919 qp->s_psn = wqe->psn;
921 if (++qp->s_last >= qp->s_size)
923 if (qp->s_last == qp->s_tail)
925 wqe = get_swqe_ptr(qp, qp->s_last);
929 switch (aeth >> 29) {
932 /* If this is a partial ACK, reset the retransmit timer. */
933 if (qp->s_last != qp->s_tail) {
934 spin_lock(&dev->pending_lock);
935 list_add_tail(&qp->timerwait,
936 &dev->pending[dev->pending_index]);
937 spin_unlock(&dev->pending_lock);
939 * If we get a partial ACK for a resent operation,
940 * we can stop resending the earlier packets and
941 * continue with the next packet the receiver wants.
943 if (ipath_cmp24(qp->s_psn, psn) <= 0) {
944 reset_psn(qp, psn + 1);
945 tasklet_hi_schedule(&qp->s_task);
947 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
948 qp->s_state = OP(SEND_LAST);
951 ipath_get_credit(qp, aeth);
952 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
953 qp->s_retry = qp->s_retry_cnt;
954 update_last_psn(qp, psn);
958 case 1: /* RNR NAK */
960 if (qp->s_last == qp->s_tail)
962 if (qp->s_rnr_retry == 0) {
963 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
966 if (qp->s_rnr_retry_cnt < 7)
969 /* The last valid PSN is the previous PSN. */
970 update_last_psn(qp, psn - 1);
972 if (wqe->wr.opcode == IB_WR_RDMA_READ)
976 (qp->s_psn - psn) & IPATH_PSN_MASK;
981 ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
982 IPATH_AETH_CREDIT_MASK];
983 ipath_insert_rnr_queue(qp);
987 if (qp->s_last == qp->s_tail)
989 /* The last valid PSN is the previous PSN. */
990 update_last_psn(qp, psn - 1);
991 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
992 IPATH_AETH_CREDIT_MASK) {
993 case 0: /* PSN sequence error */
996 * Back up to the responder's expected PSN.
997 * Note that we might get a NAK in the middle of an
998 * RDMA READ response which terminates the RDMA
1001 ipath_restart_rc(qp, psn, &wc);
1004 case 1: /* Invalid Request */
1005 wc.status = IB_WC_REM_INV_REQ_ERR;
1006 dev->n_other_naks++;
1009 case 2: /* Remote Access Error */
1010 wc.status = IB_WC_REM_ACCESS_ERR;
1011 dev->n_other_naks++;
1014 case 3: /* Remote Operation Error */
1015 wc.status = IB_WC_REM_OP_ERR;
1016 dev->n_other_naks++;
1018 wc.wr_id = wqe->wr.wr_id;
1019 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1023 wc.src_qp = qp->remote_qpn;
1025 wc.slid = qp->remote_ah_attr.dlid;
1026 wc.sl = qp->remote_ah_attr.sl;
1027 wc.dlid_path_bits = 0;
1029 ipath_sqerror_qp(qp, &wc);
1033 /* Ignore other reserved NAK error codes */
1036 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1039 default: /* 2: reserved */
1041 /* Ignore reserved NAK codes. */
1050 * ipath_rc_rcv_resp - process an incoming RC response packet
1051 * @dev: the device this packet came in on
1052 * @ohdr: the other headers for this packet
1053 * @data: the packet data
1054 * @tlen: the packet length
1055 * @qp: the QP for this packet
1056 * @opcode: the opcode for this packet
1057 * @psn: the packet sequence number for this packet
1058 * @hdrsize: the header length
1059 * @pmtu: the path MTU
1060 * @header_in_data: true if part of the header data is in the data buffer
1062 * This is called from ipath_rc_rcv() to process an incoming RC response
1063 * packet for the given QP.
1064 * Called at interrupt level.
1066 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1067 struct ipath_other_headers *ohdr,
1068 void *data, u32 tlen,
1069 struct ipath_qp *qp,
1071 u32 psn, u32 hdrsize, u32 pmtu,
1074 struct ipath_swqe *wqe;
1075 unsigned long flags;
1081 spin_lock_irqsave(&qp->s_lock, flags);
1083 /* Ignore invalid responses. */
1084 if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1087 /* Ignore duplicate responses. */
1088 diff = ipath_cmp24(psn, qp->s_last_psn);
1089 if (unlikely(diff <= 0)) {
1090 /* Update credits for "ghost" ACKs */
1091 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1092 if (!header_in_data)
1093 aeth = be32_to_cpu(ohdr->u.aeth);
1095 aeth = be32_to_cpu(((__be32 *) data)[0]);
1096 data += sizeof(__be32);
1098 if ((aeth >> 29) == 0)
1099 ipath_get_credit(qp, aeth);
1104 if (unlikely(qp->s_last == qp->s_tail))
1106 wqe = get_swqe_ptr(qp, qp->s_last);
1109 case OP(ACKNOWLEDGE):
1110 case OP(ATOMIC_ACKNOWLEDGE):
1111 case OP(RDMA_READ_RESPONSE_FIRST):
1112 if (!header_in_data)
1113 aeth = be32_to_cpu(ohdr->u.aeth);
1115 aeth = be32_to_cpu(((__be32 *) data)[0]);
1116 data += sizeof(__be32);
1118 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1121 if (!header_in_data) {
1122 __be32 *p = ohdr->u.at.atomic_ack_eth;
1124 val = ((u64) be32_to_cpu(p[0]) << 32) |
1127 val = be64_to_cpu(((__be64 *) data)[0]);
1128 *(u64 *) wqe->sg_list[0].vaddr = val;
1130 if (!do_rc_ack(qp, aeth, psn, opcode) ||
1131 opcode != OP(RDMA_READ_RESPONSE_FIRST))
1134 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1137 * If this is a response to a resent RDMA read, we
1138 * have to be careful to copy the data to the right
1141 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1145 case OP(RDMA_READ_RESPONSE_MIDDLE):
1146 /* no AETH, no ACK */
1147 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1149 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1152 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1155 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1157 if (unlikely(pmtu >= qp->s_rdma_read_len))
1160 /* We got a response so update the timeout. */
1161 spin_lock(&dev->pending_lock);
1162 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1163 list_move_tail(&qp->timerwait,
1164 &dev->pending[dev->pending_index]);
1165 spin_unlock(&dev->pending_lock);
1167 * Update the RDMA receive state but do the copy w/o
1168 * holding the locks and blocking interrupts.
1170 qp->s_rdma_read_len -= pmtu;
1171 update_last_psn(qp, psn);
1172 spin_unlock_irqrestore(&qp->s_lock, flags);
1173 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1176 case OP(RDMA_READ_RESPONSE_ONLY):
1177 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1179 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1182 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1184 /* Get the number of bytes the message was padded by. */
1185 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1187 * Check that the data size is >= 0 && <= pmtu.
1188 * Remember to account for the AETH header (4) and
1191 if (unlikely(tlen < (hdrsize + pad + 8)))
1194 * If this is a response to a resent RDMA read, we
1195 * have to be careful to copy the data to the right
1198 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1202 case OP(RDMA_READ_RESPONSE_LAST):
1203 /* ACKs READ req. */
1204 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1206 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1209 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1211 /* Get the number of bytes the message was padded by. */
1212 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1214 * Check that the data size is >= 1 && <= pmtu.
1215 * Remember to account for the AETH header (4) and
1218 if (unlikely(tlen <= (hdrsize + pad + 8)))
1221 tlen -= hdrsize + pad + 8;
1222 if (unlikely(tlen != qp->s_rdma_read_len))
1224 if (!header_in_data)
1225 aeth = be32_to_cpu(ohdr->u.aeth);
1227 aeth = be32_to_cpu(((__be32 *) data)[0]);
1228 data += sizeof(__be32);
1230 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1231 (void) do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST));
1236 spin_unlock_irqrestore(&qp->s_lock, flags);
1240 wc.status = IB_WC_LOC_QP_OP_ERR;
1244 wc.status = IB_WC_LOC_LEN_ERR;
1246 wc.wr_id = wqe->wr.wr_id;
1247 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1252 wc.src_qp = qp->remote_qpn;
1255 wc.slid = qp->remote_ah_attr.dlid;
1256 wc.sl = qp->remote_ah_attr.sl;
1257 wc.dlid_path_bits = 0;
1259 ipath_sqerror_qp(qp, &wc);
1260 spin_unlock_irqrestore(&qp->s_lock, flags);
1266 * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1267 * @dev: the device this packet came in on
1268 * @ohdr: the other headers for this packet
1269 * @data: the packet data
1270 * @qp: the QP for this packet
1271 * @opcode: the opcode for this packet
1272 * @psn: the packet sequence number for this packet
1273 * @diff: the difference between the PSN and the expected PSN
1274 * @header_in_data: true if part of the header data is in the data buffer
1276 * This is called from ipath_rc_rcv() to process an unexpected
1277 * incoming RC packet for the given QP.
1278 * Called at interrupt level.
1279 * Return 1 if no more processing is needed; otherwise return 0 to
1280 * schedule a response to be sent.
1282 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1283 struct ipath_other_headers *ohdr,
1285 struct ipath_qp *qp,
1291 struct ipath_ack_entry *e;
1294 unsigned long flags;
1298 * Packet sequence error.
1299 * A NAK will ACK earlier sends and RDMA writes.
1300 * Don't queue the NAK if we already sent one.
1302 if (!qp->r_nak_state) {
1303 qp->r_nak_state = IB_NAK_PSN_ERROR;
1304 /* Use the expected PSN. */
1305 qp->r_ack_psn = qp->r_psn;
1312 * Handle a duplicate request. Don't re-execute SEND, RDMA
1313 * write or atomic op. Don't NAK errors, just silently drop
1314 * the duplicate request. Note that r_sge, r_len, and
1315 * r_rcv_len may be in use so don't modify them.
1317 * We are supposed to ACK the earliest duplicate PSN but we
1318 * can coalesce an outstanding duplicate ACK. We have to
1319 * send the earliest so that RDMA reads can be restarted at
1320 * the requester's expected PSN.
1322 * First, find where this duplicate PSN falls within the
1323 * ACKs previously sent.
1325 psn &= IPATH_PSN_MASK;
1328 spin_lock_irqsave(&qp->s_lock, flags);
1329 for (i = qp->r_head_ack_queue; ; i = prev) {
1330 if (i == qp->s_tail_ack_queue)
1335 prev = IPATH_MAX_RDMA_ATOMIC;
1336 if (prev == qp->r_head_ack_queue) {
1340 e = &qp->s_ack_queue[prev];
1345 if (ipath_cmp24(psn, e->psn) >= 0)
1349 case OP(RDMA_READ_REQUEST): {
1350 struct ib_reth *reth;
1355 * If we didn't find the RDMA read request in the ack queue,
1356 * or the send tasklet is already backed up to send an
1357 * earlier entry, we can ignore this request.
1359 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1361 /* RETH comes after BTH */
1362 if (!header_in_data)
1363 reth = &ohdr->u.rc.reth;
1365 reth = (struct ib_reth *)data;
1366 data += sizeof(*reth);
1369 * Address range must be a subset of the original
1370 * request and start on pmtu boundaries.
1371 * We reuse the old ack_queue slot since the requester
1372 * should not back up and request an earlier PSN for the
1375 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1376 ib_mtu_enum_to_int(qp->path_mtu);
1377 len = be32_to_cpu(reth->length);
1378 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1381 u32 rkey = be32_to_cpu(reth->rkey);
1382 u64 vaddr = be64_to_cpu(reth->vaddr);
1385 ok = ipath_rkey_ok(qp, &e->rdma_sge,
1387 IB_ACCESS_REMOTE_READ);
1391 e->rdma_sge.sg_list = NULL;
1392 e->rdma_sge.num_sge = 0;
1393 e->rdma_sge.sge.mr = NULL;
1394 e->rdma_sge.sge.vaddr = NULL;
1395 e->rdma_sge.sge.length = 0;
1396 e->rdma_sge.sge.sge_length = 0;
1399 qp->s_ack_state = OP(ACKNOWLEDGE);
1400 qp->s_tail_ack_queue = prev;
1404 case OP(COMPARE_SWAP):
1405 case OP(FETCH_ADD): {
1407 * If we didn't find the atomic request in the ack queue
1408 * or the send tasklet is already backed up to send an
1409 * earlier entry, we can ignore this request.
1411 if (!e || e->opcode != (u8) opcode || old_req)
1413 qp->s_ack_state = OP(ACKNOWLEDGE);
1414 qp->s_tail_ack_queue = prev;
1422 * Resend the most recent ACK if this request is
1423 * after all the previous RDMA reads and atomics.
1425 if (i == qp->r_head_ack_queue) {
1426 spin_unlock_irqrestore(&qp->s_lock, flags);
1427 qp->r_nak_state = 0;
1428 qp->r_ack_psn = qp->r_psn - 1;
1432 * Resend the RDMA read or atomic op which
1433 * ACKs this duplicate request.
1435 qp->s_ack_state = OP(ACKNOWLEDGE);
1436 qp->s_tail_ack_queue = i;
1439 qp->r_nak_state = 0;
1440 tasklet_hi_schedule(&qp->s_task);
1443 spin_unlock_irqrestore(&qp->s_lock, flags);
1451 static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
1453 unsigned long flags;
1455 spin_lock_irqsave(&qp->s_lock, flags);
1456 qp->state = IB_QPS_ERR;
1457 ipath_error_qp(qp, err);
1458 spin_unlock_irqrestore(&qp->s_lock, flags);
1462 * ipath_rc_rcv - process an incoming RC packet
1463 * @dev: the device this packet came in on
1464 * @hdr: the header of this packet
1465 * @has_grh: true if the header has a GRH
1466 * @data: the packet data
1467 * @tlen: the packet length
1468 * @qp: the QP for this packet
1470 * This is called from ipath_qp_rcv() to process an incoming RC packet
1472 * Called at interrupt level.
1474 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1475 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1477 struct ipath_other_headers *ohdr;
1483 u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1485 struct ib_reth *reth;
1488 /* Validate the SLID. See Ch. 9.6.1.5 */
1489 if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1495 hdrsize = 8 + 12; /* LRH + BTH */
1496 psn = be32_to_cpu(ohdr->bth[2]);
1499 ohdr = &hdr->u.l.oth;
1500 hdrsize = 8 + 40 + 12; /* LRH + GRH + BTH */
1502 * The header with GRH is 60 bytes and the core driver sets
1503 * the eager header buffer size to 56 bytes so the last 4
1504 * bytes of the BTH header (PSN) is in the data buffer.
1506 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1507 if (header_in_data) {
1508 psn = be32_to_cpu(((__be32 *) data)[0]);
1509 data += sizeof(__be32);
1511 psn = be32_to_cpu(ohdr->bth[2]);
1515 * Process responses (ACKs) before anything else. Note that the
1516 * packet sequence number will be for something in the send work
1517 * queue rather than the expected receive packet sequence number.
1518 * In other words, this QP is the requester.
1520 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1521 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1522 opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1523 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1524 hdrsize, pmtu, header_in_data);
1528 /* Compute 24 bits worth of difference. */
1529 diff = ipath_cmp24(psn, qp->r_psn);
1530 if (unlikely(diff)) {
1531 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1532 psn, diff, header_in_data))
1537 /* Check for opcode sequence errors. */
1538 switch (qp->r_state) {
1539 case OP(SEND_FIRST):
1540 case OP(SEND_MIDDLE):
1541 if (opcode == OP(SEND_MIDDLE) ||
1542 opcode == OP(SEND_LAST) ||
1543 opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1546 ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
1547 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1548 qp->r_ack_psn = qp->r_psn;
1551 case OP(RDMA_WRITE_FIRST):
1552 case OP(RDMA_WRITE_MIDDLE):
1553 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1554 opcode == OP(RDMA_WRITE_LAST) ||
1555 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1560 if (opcode == OP(SEND_MIDDLE) ||
1561 opcode == OP(SEND_LAST) ||
1562 opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1563 opcode == OP(RDMA_WRITE_MIDDLE) ||
1564 opcode == OP(RDMA_WRITE_LAST) ||
1565 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1568 * Note that it is up to the requester to not send a new
1569 * RDMA read or atomic operation before receiving an ACK
1570 * for the previous operation.
1578 /* OK, process the packet. */
1580 case OP(SEND_FIRST):
1581 if (!ipath_get_rwqe(qp, 0)) {
1584 * A RNR NAK will ACK earlier sends and RDMA writes.
1585 * Don't queue the NAK if a RDMA read or atomic
1586 * is pending though.
1588 if (qp->r_nak_state)
1590 qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1591 qp->r_ack_psn = qp->r_psn;
1596 case OP(SEND_MIDDLE):
1597 case OP(RDMA_WRITE_MIDDLE):
1599 /* Check for invalid length PMTU or posted rwqe len. */
1600 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1602 qp->r_rcv_len += pmtu;
1603 if (unlikely(qp->r_rcv_len > qp->r_len))
1605 ipath_copy_sge(&qp->r_sge, data, pmtu);
1608 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1610 if (!ipath_get_rwqe(qp, 1))
1615 case OP(SEND_ONLY_WITH_IMMEDIATE):
1616 if (!ipath_get_rwqe(qp, 0))
1619 if (opcode == OP(SEND_ONLY))
1622 case OP(SEND_LAST_WITH_IMMEDIATE):
1624 if (header_in_data) {
1625 wc.imm_data = *(__be32 *) data;
1626 data += sizeof(__be32);
1628 /* Immediate data comes after BTH */
1629 wc.imm_data = ohdr->u.imm_data;
1632 wc.wc_flags = IB_WC_WITH_IMM;
1635 case OP(RDMA_WRITE_LAST):
1637 /* Get the number of bytes the message was padded by. */
1638 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1639 /* Check for invalid length. */
1640 /* XXX LAST len should be >= 1 */
1641 if (unlikely(tlen < (hdrsize + pad + 4)))
1643 /* Don't count the CRC. */
1644 tlen -= (hdrsize + pad + 4);
1645 wc.byte_len = tlen + qp->r_rcv_len;
1646 if (unlikely(wc.byte_len > qp->r_len))
1648 ipath_copy_sge(&qp->r_sge, data, tlen);
1650 if (!qp->r_wrid_valid)
1652 qp->r_wrid_valid = 0;
1653 wc.wr_id = qp->r_wr_id;
1654 wc.status = IB_WC_SUCCESS;
1655 wc.opcode = IB_WC_RECV;
1658 wc.src_qp = qp->remote_qpn;
1660 wc.slid = qp->remote_ah_attr.dlid;
1661 wc.sl = qp->remote_ah_attr.sl;
1662 wc.dlid_path_bits = 0;
1664 /* Signal completion event if the solicited bit is set. */
1665 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1667 __constant_cpu_to_be32(1 << 23)) != 0);
1670 case OP(RDMA_WRITE_FIRST):
1671 case OP(RDMA_WRITE_ONLY):
1672 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1674 /* RETH comes after BTH */
1675 if (!header_in_data)
1676 reth = &ohdr->u.rc.reth;
1678 reth = (struct ib_reth *)data;
1679 data += sizeof(*reth);
1681 hdrsize += sizeof(*reth);
1682 qp->r_len = be32_to_cpu(reth->length);
1684 if (qp->r_len != 0) {
1685 u32 rkey = be32_to_cpu(reth->rkey);
1686 u64 vaddr = be64_to_cpu(reth->vaddr);
1689 /* Check rkey & NAK */
1690 ok = ipath_rkey_ok(qp, &qp->r_sge,
1691 qp->r_len, vaddr, rkey,
1692 IB_ACCESS_REMOTE_WRITE);
1696 qp->r_sge.sg_list = NULL;
1697 qp->r_sge.sge.mr = NULL;
1698 qp->r_sge.sge.vaddr = NULL;
1699 qp->r_sge.sge.length = 0;
1700 qp->r_sge.sge.sge_length = 0;
1702 if (unlikely(!(qp->qp_access_flags &
1703 IB_ACCESS_REMOTE_WRITE)))
1705 if (opcode == OP(RDMA_WRITE_FIRST))
1707 else if (opcode == OP(RDMA_WRITE_ONLY))
1709 if (!ipath_get_rwqe(qp, 1))
1713 case OP(RDMA_READ_REQUEST): {
1714 struct ipath_ack_entry *e;
1718 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1720 next = qp->r_head_ack_queue + 1;
1721 if (next > IPATH_MAX_RDMA_ATOMIC)
1723 if (unlikely(next == qp->s_tail_ack_queue))
1725 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1726 /* RETH comes after BTH */
1727 if (!header_in_data)
1728 reth = &ohdr->u.rc.reth;
1730 reth = (struct ib_reth *)data;
1731 data += sizeof(*reth);
1733 len = be32_to_cpu(reth->length);
1735 u32 rkey = be32_to_cpu(reth->rkey);
1736 u64 vaddr = be64_to_cpu(reth->vaddr);
1739 /* Check rkey & NAK */
1740 ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1741 rkey, IB_ACCESS_REMOTE_READ);
1745 * Update the next expected PSN. We add 1 later
1746 * below, so only add the remainder here.
1749 qp->r_psn += (len - 1) / pmtu;
1751 e->rdma_sge.sg_list = NULL;
1752 e->rdma_sge.num_sge = 0;
1753 e->rdma_sge.sge.mr = NULL;
1754 e->rdma_sge.sge.vaddr = NULL;
1755 e->rdma_sge.sge.length = 0;
1756 e->rdma_sge.sge.sge_length = 0;
1761 * We need to increment the MSN here instead of when we
1762 * finish sending the result since a duplicate request would
1763 * increment it more than once.
1767 qp->r_state = opcode;
1768 qp->r_nak_state = 0;
1770 qp->r_head_ack_queue = next;
1772 /* Call ipath_do_rc_send() in another thread. */
1773 tasklet_hi_schedule(&qp->s_task);
1778 case OP(COMPARE_SWAP):
1779 case OP(FETCH_ADD): {
1780 struct ib_atomic_eth *ateth;
1781 struct ipath_ack_entry *e;
1788 if (unlikely(!(qp->qp_access_flags &
1789 IB_ACCESS_REMOTE_ATOMIC)))
1791 next = qp->r_head_ack_queue + 1;
1792 if (next > IPATH_MAX_RDMA_ATOMIC)
1794 if (unlikely(next == qp->s_tail_ack_queue))
1796 if (!header_in_data)
1797 ateth = &ohdr->u.atomic_eth;
1799 ateth = (struct ib_atomic_eth *)data;
1800 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1801 be32_to_cpu(ateth->vaddr[1]);
1802 if (unlikely(vaddr & (sizeof(u64) - 1)))
1804 rkey = be32_to_cpu(ateth->rkey);
1805 /* Check rkey & NAK */
1806 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1807 sizeof(u64), vaddr, rkey,
1808 IB_ACCESS_REMOTE_ATOMIC)))
1810 /* Perform atomic OP and save result. */
1811 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1812 sdata = be64_to_cpu(ateth->swap_data);
1813 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1814 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1815 (u64) atomic64_add_return(sdata, maddr) - sdata :
1816 (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1817 be64_to_cpu(ateth->compare_data),
1820 e->psn = psn & IPATH_PSN_MASK;
1823 qp->r_state = opcode;
1824 qp->r_nak_state = 0;
1826 qp->r_head_ack_queue = next;
1828 /* Call ipath_do_rc_send() in another thread. */
1829 tasklet_hi_schedule(&qp->s_task);
1835 /* NAK unknown opcodes. */
1839 qp->r_state = opcode;
1840 qp->r_ack_psn = psn;
1841 qp->r_nak_state = 0;
1842 /* Send an ACK if requested or required. */
1843 if (psn & (1 << 31))
1848 ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1849 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1850 qp->r_ack_psn = qp->r_psn;